Database Migration Plan

Phase 1 — EXPAND
  Deploy migration: add new column (nullable), create new indexes
  Old app: continues to work (ignores new column)
  New app: not yet deployed
  Duration: [~X min] | Rollback: trivial — drop new column

       │
       ▼

Phase 2 — BACKFILL + DUAL-WRITE
  Deploy app update: writes to both old and new columns
  Run backfill: populate new column for existing rows
  Validate: confirm 100% of rows have non-null new column
  Duration: [~X hours depending on data volume]
  Rollback: deploy previous app version; new column is still nullable

       │
       ▼

Phase 3 — ENFORCE + SWITCH
  Deploy migration: add NOT NULL constraint, drop old column/index
  Deploy app update: reads only from new column
  Duration: [~X min] | Rollback: requires forward-fix (constraint must be dropped first)

       │
       ▼

Phase 4 — CONTRACT (optional cleanup)
  Deploy migration: drop deprecated columns, rename if needed
  Final state matches target schema
  Rollback: not recommended — contract changes are destructive

-- Migration: 001_add_org_id_to_users.sql
BEGIN;

-- Add nullable column (safe — old app ignores it)
ALTER TABLE users
    ADD COLUMN org_id UUID NULL
        REFERENCES organisations(id) ON DELETE RESTRICT;

-- Add index NOW, not in Phase 3 — building index on large table during Phase 3 is risky
CREATE INDEX CONCURRENTLY users_org_id_idx ON users (org_id);

-- Note: CONCURRENTLY does not lock the table; safe on live traffic
-- Note: Cannot run CONCURRENTLY inside a transaction block; run separately if needed

COMMIT;

-- Confirm column exists and is nullable
SELECT column_name, data_type, is_nullable
FROM information_schema.columns
WHERE table_name = 'users' AND column_name = 'org_id';
-- Expected: is_nullable = 'YES'

-- Confirm index exists
SELECT indexname, indexdef
FROM pg_indexes
WHERE tablename = 'users' AND indexname = 'users_org_id_idx';

BEGIN;
DROP INDEX CONCURRENTLY IF EXISTS users_org_id_idx;
ALTER TABLE users DROP COLUMN IF EXISTS org_id;
COMMIT;

// All INSERT and UPDATE operations must now set BOTH old_column and new_column
// until Phase 3 is complete. This ensures new rows are populated during the backfill window.

-- Run in batches to avoid locking. Adjust batch size based on table size and DB load.
-- Target: no single batch takes more than 5 seconds.

DO $$
DECLARE
    batch_size  INT := 1000;
    affected    INT;
BEGIN
    LOOP
        UPDATE users
        SET    org_id = accounts.organisation_id
        FROM   accounts
        WHERE  users.account_id = accounts.id
          AND  users.org_id IS NULL
        LIMIT  batch_size;

        GET DIAGNOSTICS affected = ROW_COUNT;
        EXIT WHEN affected = 0;

        -- Pause between batches to avoid saturating I/O
        PERFORM pg_sleep(0.1);
    END LOOP;
END $$;

-- Check progress — run periodically during backfill
SELECT
    COUNT(*) FILTER (WHERE org_id IS NOT NULL) AS backfilled,
    COUNT(*) FILTER (WHERE org_id IS NULL)     AS remaining,
    COUNT(*)                                   AS total,
    ROUND(
        100.0 * COUNT(*) FILTER (WHERE org_id IS NOT NULL) / COUNT(*), 2
    ) AS pct_complete
FROM users;

-- Must return 0 before proceeding to Phase 3
SELECT COUNT(*) AS unbackfilled_rows
FROM users
WHERE org_id IS NULL;

-- Confirm no new rows written without org_id (dual-write working)
SELECT COUNT(*) AS recent_missing
FROM users
WHERE org_id IS NULL
  AND created_at > now() - INTERVAL '1 hour';

-- Step 1: Add constraint as NOT VALID (no full table scan — instant)
ALTER TABLE users
    ADD CONSTRAINT users_org_id_not_null
    CHECK (org_id IS NOT NULL) NOT VALID;

-- Step 2: VALIDATE CONSTRAINT (takes a SHARE UPDATE EXCLUSIVE lock — allows reads and writes)
-- Run this separately, as it can take minutes on large tables
ALTER TABLE users
    VALIDATE CONSTRAINT users_org_id_not_null;

-- Step 3: Once validated, convert to actual NOT NULL
-- (PostgreSQL trusts the validated check constraint — this is instant)
ALTER TABLE users
    ALTER COLUMN org_id SET NOT NULL;

-- Step 4: Drop the now-redundant check constraint
ALTER TABLE users
    DROP CONSTRAINT users_org_id_not_null;

-- Confirm NOT NULL is enforced
SELECT column_name, is_nullable
FROM information_schema.columns
WHERE table_name = 'users' AND column_name = 'org_id';
-- Expected: is_nullable = 'NO'

-- Test that insert without org_id fails (run in a transaction and roll back)
BEGIN;
INSERT INTO users (email) VALUES ('test@example.com');
-- Expected: ERROR: null value in column "org_id" violates not-null constraint
ROLLBACK;

-- Drop the NOT NULL constraint (restores nullable state)
ALTER TABLE users ALTER COLUMN org_id DROP NOT NULL;
-- Then deploy previous app version (dual-write)
-- Note: Once app code reading the new column is live, rolling back the constraint
-- without rolling back the app will cause issues — plan this carefully.

BEGIN;

-- Drop the old column
ALTER TABLE users DROP COLUMN account_id;

-- Drop any indexes that referenced the old column
DROP INDEX IF EXISTS users_account_id_idx;

COMMIT;

-- Confirm no application queries still reference the old column
-- (Check this in code review and via a search of the codebase before running)
-- grep -r "account_id" app/

-- Confirm the column is safe to drop
SELECT COUNT(*) FROM users WHERE account_id IS NOT NULL;
-- Should be 0 (or irrelevant once new column is canonical)

-- Record these values before any migration step
SELECT COUNT(*)   AS total_users FROM users;
SELECT COUNT(*)   AS total_orgs  FROM organisations;
SELECT MIN(created_at), MAX(created_at) FROM users;

-- Check for any anomalies in the source data before backfill
SELECT COUNT(*) AS users_without_account
FROM users WHERE account_id IS NULL;

-- All users have an org that exists
SELECT COUNT(*) AS orphaned_org_refs
FROM users u
WHERE u.org_id IS NOT NULL
  AND NOT EXISTS (
      SELECT 1 FROM organisations o WHERE o.id = u.org_id
  );
-- Expected: 0

-- org_id matches expected value from source column
SELECT COUNT(*) AS mismatched_backfill
FROM users u
JOIN accounts a ON u.account_id = a.id
WHERE u.org_id != a.organisation_id;
-- Expected: 0

-- Row count unchanged (no rows created or deleted by migration)
SELECT COUNT(*) AS total_users_after FROM users;
-- Must match pre-migration baseline

-- Old column is gone
SELECT COUNT(*) FROM information_schema.columns
WHERE table_name = 'users' AND column_name = 'account_id';
-- Expected: 0

-- New column is NOT NULL
SELECT is_nullable FROM information_schema.columns
WHERE table_name = 'users' AND column_name = 'org_id';
-- Expected: NO